The Third Field

In January 2026, Nielsen Norman Group published an independent evaluation of AI interviewers, conducted in September 2025. One participant described the experience afterward: “It did a good job of […] mirroring back to me what I had said so that I felt […] heard and understood […] that felt reflective of what a real conversation feels like with another human.” The same evaluation also counted the seams — interruptions, awkward pauses, praise where none was earned — and concluded that the interviewers it tested belong in structured interviews, not open-ended discovery. Both findings are accurate, and this essay takes both seriously.

That an AI can hold an interview a participant experiences as a real conversation is, by now, the least interesting part of the story. The interesting part is what happens to research when interviewing stops being scarce. Because once every respondent in a structured study can also be asked why — in conversation, with follow-ups, probed when the first answer is thin — the deliverable changes shape. A finding is no longer a percentage with a bag of coded fragments beside it. It is a percentage a skeptical stakeholder can interrogate: click from the number down to the reasoning of the exact people who produced it, in their own words, at the full sample size.

That sentence — a percentage you can interrogate like an interview — is not something either research discipline can produce. A survey delivers the number and an open-text box full of fragments someone codes into themes. A qualitative study delivers the why, vividly, from twelve people, and the stakeholder asks whether twelve is enough to bet a launch on. For the whole history of the field, you chose: the number or the why. The choice felt like methodology. It was never methodology. It was logistics.

Here is the logistics, in arithmetic. A trained moderator can hold one conversation at a time. Six to eight in a good week, once recruiting, scheduling, and the analysis that trails each session are counted in. A thirty-interview study therefore takes a quarter and costs what a small car costs — so qualitative samples are small, not because depth demands smallness, but because one human can only do so much. Point the same constraint the other way and you get the survey: if the answer must come from a thousand people and no moderator can sit with each of them, the instrument has to run itself, and a self-running instrument cannot ask the obvious next question. Closed-ended questions are not an epistemological commitment. They are what you write when nobody will be in the room.

Everything the industry treats as a law of nature — sample-size conventions, three-month timelines, the wall between “qual” and “quant” with its two professional associations, two conference circuits, and two career tracks — grew out of that single operational fact. The fact has now lapsed. An AI agent conducts interviews in parallel, in the participant’s language, with the patience to probe the two-hundredth conversation as carefully as the first. What that unlocks is not cheaper versions of the studies you already run. It is a third kind of study, with its own logic, and it needs its own name before procurement gives it the wrong one.

Two things frame everything that follows. The first is that this shift does not need anyone’s permission. The economics of conducting a conversation have changed by orders of magnitude, and no industry has held a price umbrella over a collapsed cost structure for long. The question is not whether these methods arrive — they are already in the building, often through a side door, run by a strategy team in a chatbot — but who will learn to use them well. The second is that nobody has, fully. Not the vendors, not the early adopters, not the professional bodies. The protocols, quality standards, and pricing logics are being worked out right now by practitioners comparing notes, and this essay is a contribution to that — not a verdict. I write it from inside the shift, as co-founder of Userflix, a firm that builds the infrastructure it describes. Discount accordingly. But every claim here rests on studies that were designed, fielded, paid for, and argued over — including the ones that went wrong — and where a limit is real I will say so plainly, because the fastest way to lose a researcher’s trust is to pretend the method has none.

Sort the questions customer research is actually asked to answer and they fall along two axes. One is what is being asked: the behavioral surface — what people do, buy, choose — or the interior, why they do it, how the decision actually ran. The other is how many people the question is about: a handful, understood in detail, or enough to generalize from.

Two axes, four cells. What about many is the survey’s home turf, and the industry is superb at it. Why about few belongs to the depth interview and the focus group — superb again. The other two cells have sat empty for forty years, not because nobody wanted them, but because no method could reach them.

Why about many is the cell everyone wanted. It is the reason every survey carries open-text boxes, and the reason those boxes are the worst part of every survey: the respondent types a phrase, an analyst who never met them codes the phrase, and the resulting “driver analysis” gestures at reasoning without ever containing any. It is also the reason agencies stitch hybrid studies together — a quant phase, then a qual phase on a subsample, double the cost and half the coherence, with the two datasets never quite answering to each other.

What about few — granular, comparable behavioral detail across individual cases — got diary studies and shop-alongs, each compromised by the same bottleneck as everything else.

The industry’s response to the two empty cells was to stop proposing studies that lived there. This is the quietest cost of a long-standing constraint: the briefs shape themselves to the available methods before anyone writes them down, and after enough years the forbidden questions are simply forgotten. Ask an insights team what they would study if interviews were as scalable as surveys and you get a strange pause — then a list. The cross-market study of how a category is actually bought, in the buyer’s own language, at depth, in five countries. The segmentation grounded in how customers reason rather than how they score attitude batteries. The tracker that explains its own movements instead of triggering a six-week root-cause study every time a number twitches.

Those studies are fieldable now. A two-hundred-person conversational study — real interviews, probed follow-ups, fielded in the respondents’ native language with reporting in English — runs on a quant-study timeline: days of fieldwork, gated by recruitment pace rather than moderator availability. The raw material is two hundred people explaining themselves. When an insights lead sees such a study delivered for the first time, the reaction is rarely about efficiency. It is the pause, and then the list — the questions they trained themselves to stop asking.

The two empty cells are where the third field lives. It has two working faces. Deep quant keeps the structured instrument — the scales, the choice tasks, the tracker spine — and threads conversation through it at every question where the why is load-bearing. Wide qual inverts the emphasis: the interview carries the study, and structure ensures that eight hundred conversations cover the same ground well enough to compare. The rest of this essay is about what those two look like in practice, where they break, and why the hardest problems sit not in the interviewing but in the analysis — which is exactly where the industry’s skepticism, for once, is pointed at the right target.

Deep quant is the easier one to picture, because it upgrades studies you already run. The instrument keeps its quantitative shape — the scales, the rankings, the choice tasks, the tracker spine. What changes is that at the questions where the why is load-bearing, the instrument opens into conversation. A respondent rates a concept four out of ten, and instead of moving on, the interviewer asks what would make it a seven. The respondent explains; the interviewer follows up once or twice, until the reasoning is actually on the table rather than gestured at; the session returns to the structured flow. The respondent experiences one coherent conversation, not a survey with an interview stapled to it — and because the conversational layer runs inside each session rather than as a second fieldwork phase, the timeline does not move.

The craft is knowing where to probe. Probe everywhere and the instrument bloats, the respondent fatigues, the analysis drowns; probe nowhere and you have built the old survey with extra steps. Three questions carry most of the decision: does this answer drive the decision the study informs, is the closed-ended response ambiguous on its face, and will this number move over time — because a tracker that moves without reasoning attached is precisely the thing that triggers the six-week root-cause study the upgrade is meant to retire. In our own fielding of studies built this way, for brands making concrete portfolio decisions, the consistent pattern is that the numbers alone would have supported two or three competing interpretations, and the conversational layer eliminated all but one. The value is not a richer report. It is that the meeting where the report is discussed ends with a decision instead of a request for another study.

Wide qual is the face with no precedent to upgrade: conversational research at sample sizes that have always belonged to surveys. Several hundred people, each interviewed in their own language, each probed with a senior moderator’s patience, all in parallel. The studies that live here are the ones from the forgotten list — the reasoning-grounded segmentation, the cross-market inquiry into how a category is actually bought, the decision-process study where the deliverable is not a rate but a structure of reasoning, sized across a population.

The central artifact is the discussion guide, and it does different work here than it ever did in a focus-group room. It must be structured enough that every one of eight hundred conversations covers the same ground — otherwise you have eight hundred anecdotes with no basis for comparison — and loose enough that the interviewer follows each respondent’s reasoning where it goes. That balance is where one firm’s practice will separate from another’s, and it is settled empirically, not in the abstract: field a first batch, read what came back, notice where the probing was too shallow or the guide too rigid, adjust, field the next. The loop takes days. When it converges, the study fields at scale, and what comes out is a dataset with a property neither parent discipline has ever produced: patterns of reasoning that can be counted, compared across segments and markets, and traced back to the voices that produced them.

One study shape is worth singling out because it shows how far the logic reaches: expert interviews. Organizations sit on reasoning they cannot scale — the senior underwriter’s, the veteran service agent’s, the buyer who has watched a category for twenty years — and the bottleneck to capturing it has never been the analysis. It has been staffing two hundred structured expert interviews. That bottleneck is gone, and “interview every expert, not a sample of eight” is a sentence organizations have not yet learned they can say.

There is a further capability that belongs to neither face and extends both: the agent is no longer confined to a call. In messaging-based studies that run over days, respondents answer between meetings and return with a thought the next morning; they send photos of the product on the kitchen counter and voice notes recorded in the car — material that comes from inside the life rather than from an appointment about the life. And through a phone camera or a shared screen, the agent can see what the respondent sees and moderate on it live: the shelf at the moment of choice, the onboarding screen at the moment of confusion, witnessed as it happens rather than reconstructed a week later in a recall interview. This is the map’s other empty cell — granular behavioral evidence, case by case, structured enough to compare. Diary studies and shop-alongs always reached toward it. A moderator who can be present in fifty kitchens at seven in the morning actually occupies it.

A method earns trust by naming its own limits before its skeptics do, so here are ours — the ones we hold as positions, and the ones we learned by fielding.

Some research acts have meaning because a human being is in the room. Interviews with the bereaved or the traumatized, ethnography of vulnerable communities, disclosure that happens because the respondent decided to trust a particular person — the agent can extend such work at its edges, but it cannot be the instrument of the central act, and a firm that claims otherwise should worry you. The practical line is finer-grained than “sensitive sectors,” though. One insights lead at a health charity drew it precisely: interviewing patients about their illness was off the table, but donors about their motivation to give, or clinicians about their working context, was not merely acceptable — it was the right entry point. The triage is not by industry. It is by what the conversation asks of the person in it.

Some questions do not need the third field at all. A single-metric tracker whose job is a number that moves does not want probing on every wave; the probe is never free, costing respondent minutes and analyst attention. Conjoint and pricing studies elicit trade-offs through structured choices, and the trade-offs are the finding — most gain nothing from a conversational companion. And when the study is six category gatekeepers in a specialized B2B market, those are six consequential conversations, each partly a relationship; the third field has no superior offer to make there. The discipline of the new category is knowing when not to use it, and a practitioner who claims it for everything is selling, not advising.

Some limits are operational and worth stating plainly. A conversational instrument trades exact-phrasing control for depth: if wave-on-wave comparability of precise wording is load-bearing, keep that question closed-ended. Probing quality is not symmetric out of the box — an interviewer that happily explores enthusiasm must be deliberately calibrated to dig into criticism with the same persistence, and we treat that as a standing calibration target rather than a solved problem. And there is a trust ritual the new method has not yet replaced: clients like to watch fieldwork happen — the viewing room behind the mirror — and a transcript delivered afterward, however complete, is not the same experience. Anyone deploying these methods with stakeholders should plan for that gap honestly.

And some limits we simply hit. Our early pilots produced failures I would rather list myself than have quoted at me: an interviewer that repeated a question it should have dropped, a voice that did not hold stable through a session. Each was found, understood, and engineered out — that is what the calibration loop is for, and it is why the first internal test interview of any study matters more than any feature. But the lesson generalizes beyond us: in this category, fieldwork quality is not a property of the technology in general. It is a property of the protocols wrapped around it, and buyers should evaluate exactly that.

Every conversation we have with an agency or an insights team eventually arrives at the same handful of fears, usually unstated and steering everything. They deserve direct answers, because some are better founded than the industry admits and others are aimed at the wrong threat entirely.

“This replaces us.” The fear underneath all the others, and the one most worth examining. What the agent replaces is the fieldwork bottleneck — the scheduling, the moderating hours, the transcription, the first-pass coding. The researcher's work shifts to the two ends of the process: at the front, deciding what to investigate — which question, asked of whom, to inform which decision — and at the back, working out what the results mean for your context, up to and including the finding the client did not want. In between, the agent does the research. Both ends become more valuable when the middle gets fast and cheap, not less — because they become the entire offer. The researcher’s hours move up the value chain or they disappear — but that choice belongs to the researcher, not the technology. The honest version of this fear is not “the agent will take my job.” It is “my job will require different skills than the ones I was promoted for,” and that version is true, and the only answer to it is learning — which everyone in this field, ourselves included, is doing in public right now.

“It replaces us” has a second, sharper form for agencies specifically: “our clients will just do it themselves.” This fear is better founded, and adoption hesitancy makes it worse, not better. In our conversations across the European market, a consistent pattern is that end clients are a step or two ahead of their agencies — small teams that could never afford traditional qual are already running conversational studies on their own, and insights departments are piloting tools while their agencies deliberate. A client who learns the new methods from someone else has learned they need you less. The agency’s defensible ground was never fieldwork logistics; it is judgment, design craft, and the ability to stand behind a finding in a hostile room. But that ground only stays defensible if the agency is visibly fluent in the new instruments. Waiting for the methods to mature is how the ground gets ceded.

“It will eat our margins.” Only if it is sold as a discount. The gravitational pull in every procurement conversation is to file the new methods as cheaper qual — same study, fewer moderator hours, lower invoice — and any agency that accepts that framing has priced its own offer into commodity territory, competing against everyone who bought the same software. The defensible position is the one this essay has been making: these are studies that could not be run before, answering questions the client had learned not to ask, and new capability is priced as capability. That is the difference between selling the labor the agent replaced and selling the judgment it cannot replace, and the margin lives entirely in the second sentence — protected by nothing except the agency’s own insistence on it. And the arithmetic has a second side the fear usually ignores: automated fieldwork does not just lower prices, it lowers costs — the most labor-intensive, lowest-margin phase of the project disappears while the billable ends remain. The same senior team delivers more studies, and every existing engagement gains extensions that used to die at the budget line: the tracker upgraded to deep quant, the follow-up study for every unresolved why, the three additional markets that once tripled the price. Sold correctly, automation is not a margin killer — it is the first margin lever fieldwork has ever offered this industry.

“One bad AI interview in front of a client ends the relationship.” Legitimate, and the mitigation is procedural, not rhetorical. New instruments have always required validation before deployment — no one fields a questionnaire without piloting it, and no one should field an agent without internal test interviews, a calibration batch, and a first study chosen so that its stakes match its novelty. The reputational risk is real — early pilots across the industry, ours included, produced real failures when the technology was still at the start of its development. It is managed the way instrument risk has always been managed: test before you field, start where the cost of a stumble is survivable, and scale with evidence.

“Clients will hear ‘AI’ and doubt the rigor.” Some will, at first. The answer is not to hide the method — concealment is both wrong and, eventually, discovered — but to make a stronger auditability claim than the old debrief ever could, which asked the client to trust the researcher’s account of thirty conversations they never saw. Disclosed and traceable beats undisclosed and vouched-for, in front of exactly the stakeholders whose trust matters most.

“Where does the next generation learn the craft?” The quietest fear and maybe the deepest. Juniors learned research by doing fieldwork; if the agent does the fieldwork, the apprenticeship model loses its bottom rung. This one we will not pretend to have solved. Something replaces it — juniors reading calibration batches and fieldwork returns at volumes no trainee ever saw, learning design earlier because design is now the job — but the honest answer is that the profession is improvising its new training pathways in real time, and anyone who claims otherwise is selling something. It belongs on the industry’s shared agenda, alongside the standards.

What all of these fears have in common is that none is answered by waiting. The methods are learnable now, the standards are being written now, and the practitioners at the table are the ones doing early, imperfect, honest work — not the ones who postponed engagement until someone else had made it safe.

Ask researchers who have actually piloted AI-moderated studies where the disappointment sits, and it is rarely the interviews. Participants show up, talk, and often say afterward that they felt heard. The frustration concentrates one layer up: the analysis reads as summary rather than insight, quotes surface without provenance or, worse, attributed to the wrong person, themes arrive pre-flattened into the categories the model found convenient. The industry’s skepticism, aimed here, is aimed at the right target — and it deserves a specific answer, not reassurance.

The answer is architectural, not cosmetic. Hand forty transcripts to a generic language model and you get exactly what the skeptics describe: a plausible, flat summary with no accountability. An analysis layer that deserves trust is built the way a research department is staffed. In our own practice, an orchestrator runs the analysis and dispatches subagents that each work through a single transcript; their per-case findings are then merged and tested across the sample. Content and arithmetic are separated by construction: language models do the interpretive work — themes, reasoning, what a respondent actually meant — while code does the counting, so that a claim like “this concern appeared in a third of interviews” comes from counted codes, not from a model’s impression of frequency. And every claim carries its provenance: the reader can drill from any finding to the verbatim words of the respondents who ground it. The result is not a first draft waiting for a rewrite. It is a report that stands as a deliverable — themes, distributions, real voices, traceable throughout — and the first time a skeptical executive clicks from a finding to a customer’s actual voice, the methodological argument is over.

What remains for the researcher — and it is the more senior half of the work — begins where the report ends. The report says what was found; it cannot know what the findings mean for this company, in this market, against this quarter’s decision. Reading the distributions against the strategic situation, spotting the rare pattern that outweighs the frequent one, turning findings into recommendations, and embedding all of it in the business context no agent has access to: that is the human layer, and it is what clients are actually buying. The boundary, held correctly, reads: the agent delivers the findings, the researcher delivers what they mean. Skip the human layer and you hand the client findings without consequences. Distrust the agent and redo the analysis by hand, and you have rebuilt the old labor intensity with new tools. Held right, the boundary does what the industry has wanted for decades: it moves the researcher’s hours from transcription and coding to judgment and counsel — where they were always supposed to be.

Categories are defined by their standards as much as their methods, and the standards for this one are being set now, in practice, by whoever shows up with positions. Ours are these.

Fieldwork conducted by an agent should say so in the deliverable, where the reader cannot miss it. An offer that honest disclosure would undermine was mispositioned to begin with.

Real respondents and synthetic ones are different products and the line between them must be bright. Everything described in this essay involves consented, compensated human beings; a “study” generated by a model role-playing personas is a different artifact with narrower legitimate uses — pretesting a guide, drafting hypotheses — and presenting model output as customer evidence should be disqualifying. Every honest actor benefits from that line being unmissable.

Data stewardship, in this category, is not compliance overhead; it is method. People disclose more in conversation than they ever typed into a survey box, which raises the obligation, not just the risk. The positions that follow: explicit consent about what is captured and kept, retention held to the minimum the analysis requires, and research data never used to train the models that conduct the research. Procurement and data-protection officers are often described as adoption blockers. They are better understood as the first serious quality reviewers of the new category, and a vendor or agency that cannot answer them in writing has not finished building the method.

And the participant’s side of the contract matters more at scale, not less. A person who gives thirty minutes of reasoning does so on the implicit promise that someone will hear it. The third field can honor that promise better than the binary ever could — not a summary of what respondents were like, but a deliverable through which they are, literally, heard.

The moderator bottleneck shaped everything: the prices, the timelines, the two associations, the wall between the number and the why. It is gone, and no amount of professional nostalgia reinstates it. What comes next is not settled — the methods are young, the standards are drafts, and everyone working in this territory, including the people building its infrastructure, is learning it study by study. That is not a caveat. It is the invitation. What remains is the pause and the list — the questions your organization trained itself to stop asking because no method could carry them. They are askable now. Pick one: a question that matters to a named stakeholder, attached to a real decision, with stakes modest enough to forgive a young method’s stumbles — and let the deliverable make the argument no essay can. The people asking these questions early, carefully, and out loud are the ones the rest of the field will end up learning from.

Bruno Recht is the co-founder of Userflix, which builds agentic research infrastructure for agencies and in-house insights teams, and teaches AI Design at Elisava in Barcelona. This essay is meant to start an exchange, not to end one. Agreement, disagreement, experience from first studies, half-formed thoughts — all welcome: bruno@getuserflix.com. I answer every email.